Paper claims SGHM to be 1.7x faster than MODNet, but testing locally SGHM is 2x slower than MODNet

cxgincsu / SemanticGuidedHumanMatting

Robust Human Matting via Semantic Guidance, ACCV 2022.

MIT License

235 stars 36 forks source link

Paper claims SGHM to be 1.7x faster than MODNet, but testing locally SGHM is 2x slower than MODNet #20

Open ani-mal opened 6 months ago

ani-mal commented 6 months ago

Description

When testing the inference performance of MODNet and compare to SGHM, the inference for SGHM is 2x slower than MODNet as opposed to 1.7x faster as claimed in the paper.

How can I reproduce the 1.7x speedup mentioned in the paper? I benchmarked this against CUDAExecutionProvider